Abstract: In urban tunnel fires with natural ventilation shafts, complex phenomena such as multi-dimensional flow fields and variable-mass flow pose significant challenges for real-time prediction. To address this, this paper proposed a fire development prediction model based on a coupled Long Short-Term Memory (LSTM) network and 2D U-Net. Firstly, a full-scale numerical model of tunnel fire was constructed using fire dynamics simulation software, considering various fire scenarios including different heat release rates, shaft heights, and ambient wind speeds. A multi-source database comprising ceiling temperature measurements and longitudinal temperature slices was established. Second, the model extracts temporal features using LSTM, integrates environmental parameters through a multilayer perceptron (MLP), and employs feature-wise linear modulation (FiLM) to modulate the U-Net bottleneck layer, thereby achieving spatiotemporal prediction of the temperature field. To address the spatial imbalance caused by the small proportion of high-temperature regions near the fire source, a composite loss function was constructed, incorporating fluid mask, multi-task adaptive weighting, total variation, and spectral constraints. The results show that the predicted values are in good agreement with the simulation results under the test cases, with a relative error below 9% and a mean absolute error of 3.8–9.6 °C. The model captures complex flow features such as smoke backflow and vortex contours, performs stably under different prediction lead times (10–30 s), and exhibits good generalization ability for unseen operating conditions. Furthermore, the current model outperforms traditional models such as convolutional long short-term memory (ConvLSTM), temporal convolutional networks (TCN), and Transformer in both prediction accuracy and computational efficiency. This study provides an intelligent technical reference for real-time situational awareness and emergency decision-making in urban tunnel fires with natural ventilation shafts.